Skip to content

models: fix GDN normalization from max to rsqrt #28068

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
danielhanchen wants to merge 2 commits into
base: master
Choose a base branch
from
+27 −15
Open

models: fix GDN normalization from max to rsqrt #28068

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c03dc68
models: use flash-linear-attention's l2norm for gated delta net q/k
danielhanchen Aug 31, 2026
757abc9
Update src/models/models.h
danielhanchen Aug 31, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions src/models/bailingmoe3.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -281,8 +281,8 @@ llama_model_bailingmoe3::graph::graph(const llama_model & model, const llm_graph
ggml_tensor * beta = ggml_mul_mat(ctx0, layer.ssm_beta, cur);
beta = ggml_sigmoid(ctx0, ggml_reshape_4d(ctx0, beta, 1, n_head, n_seq_tokens, n_seqs));

q = ggml_l2_norm(ctx0, q, hparams.f_norm_rms_eps);
k = ggml_l2_norm(ctx0, k, hparams.f_norm_rms_eps);
q = build_gdn_l2_norm(ctx0, q, hparams.f_norm_rms_eps);
k = build_gdn_l2_norm(ctx0, k, hparams.f_norm_rms_eps);

ggml_tensor * states_all = mctx_cur->get_s_l(il);
ggml_tensor * state = build_rs(inp_rs, states_all, hparams.n_embd_s(), n_seqs);
Expand Down
6 changes: 3 additions & 3 deletions src/models/kimi-k3.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -441,9 +441,9 @@ ggml_tensor * llama_model_kimi_k3::graph::build_kda_layer(
ggml_tensor * state = build_rs(inp_rs, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_dim, head_dim, n_head_kda, n_seqs);

const float eps = hparams.f_norm_rms_eps;
Qcur = ggml_l2_norm(ctx0, Qcur, eps);
Kcur = ggml_l2_norm(ctx0, Kcur, eps);
const float eps_norm = hparams.f_norm_rms_eps;
Qcur = build_gdn_l2_norm(ctx0, Qcur, eps_norm);
Kcur = build_gdn_l2_norm(ctx0, Kcur, eps_norm);

auto attn_out = build_delta_net(Qcur, Kcur, Vcur, g1, beta, state, il);

Expand Down
5 changes: 3 additions & 2 deletions src/models/kimi-linear.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -331,10 +331,11 @@ llama_model_kimi_linear::graph::graph(const llama_model & model, const llm_graph
ggml_tensor * state = build_rs(inp_rs, ssm_states_all, hparams.n_embd_s(), n_seqs);
state = ggml_reshape_4d(ctx0, state, head_dim, head_dim, n_head, n_seqs);


const float eps_norm = hparams.f_norm_rms_eps;

Qcur = ggml_l2_norm(ctx0, Qcur, eps_norm);
Kcur = ggml_l2_norm(ctx0, Kcur, eps_norm);
Qcur = build_gdn_l2_norm(ctx0, Qcur, eps_norm);
Kcur = build_gdn_l2_norm(ctx0, Kcur, eps_norm);

// Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens
auto attn_out = build_delta_net(Qcur, Kcur, Vcur, g1, beta, state, il);
Expand Down
7 changes: 7 additions & 0 deletions src/models/models.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,13 @@

class llama_memory_hybrid_idx_context;

// ref: https://github.com/ggml-org/llama.cpp/pull/28068
static inline ggml_tensor * build_gdn_l2_norm(ggml_context * ctx, ggml_tensor * x, float eps) {
const float n = x->ne[0];

return ggml_scale(ctx, ggml_rms_norm(ctx, x, eps/n), 1.0f/sqrtf(n));
}

Comment thread
danielhanchen marked this conversation as resolved.
//
// base classes
//
Expand Down
5 changes: 3 additions & 2 deletions src/models/qwen35.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -427,10 +427,11 @@ ggml_tensor * llama_model_qwen35::graph::build_layer_attn_linear(
cb(k_conv, "k_conv", il);
cb(v_conv, "v_conv", il);


const float eps_norm = hparams.f_norm_rms_eps;

q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);
k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);
q_conv = build_gdn_l2_norm(ctx0, q_conv, eps_norm);
k_conv = build_gdn_l2_norm(ctx0, k_conv, eps_norm);

//q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
//k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
Expand Down
5 changes: 3 additions & 2 deletions src/models/qwen35moe.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -451,10 +451,11 @@ ggml_tensor * llama_model_qwen35moe::graph::build_layer_attn_linear(
cb(k_conv, "k_conv", il);
cb(v_conv, "v_conv", il);


const float eps_norm = hparams.f_norm_rms_eps;

q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);
k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);
q_conv = build_gdn_l2_norm(ctx0, q_conv, eps_norm);
k_conv = build_gdn_l2_norm(ctx0, k_conv, eps_norm);

//q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
//k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
Expand Down
5 changes: 3 additions & 2 deletions src/models/qwen3next.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -507,10 +507,11 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_attn_linear(
cb(k_conv, "k_conv", il);
cb(v_conv, "v_conv", il);


const float eps_norm = hparams.f_norm_rms_eps;

q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);
k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);
q_conv = build_gdn_l2_norm(ctx0, q_conv, eps_norm);
k_conv = build_gdn_l2_norm(ctx0, k_conv, eps_norm);

//q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
//k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
Expand Down
5 changes: 3 additions & 2 deletions src/models/qwen4exp.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -875,10 +875,11 @@ ggml_tensor * llama_model_qwen4exp::graph::build_layer_attn_linear(
cb(k_conv, "k_conv", il);
cb(v_conv, "v_conv", il);


const float eps_norm = hparams.f_norm_rms_eps;

q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);
k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);
q_conv = build_gdn_l2_norm(ctx0, q_conv, eps_norm);
k_conv = build_gdn_l2_norm(ctx0, k_conv, eps_norm);

// repeat to match shapes when head keys != value keys; unneeded with the fused GDN
if (num_k_heads != num_v_heads && (!cparams.fused_gdn_ar || !cparams.fused_gdn_ch)) {
Expand Down
Loading